We investigate in this paper the adequate unit of analysis for Arabic Mention Detection. We experiment different segmentation schemes with various feature-sets. Results show that when limited resources are available, models built on morphologically segmented data outperform other models by up to 4F points. On the other hand, when more resources extracted from morphologically segmented data become available, models built with Arabic TreeBank style segmentation yield to better results. We also show additional improvement by combining different segmentation schemes. 1